train-model-details.md

Trained Models

Baseline

• full swiss text data
• german text corpus kenlm
• learning rate 0.0005
• dropout 0.4
• epochs ca. 60
• n-hiddenstates 1024
• tested on test07.csv
• WER: 0.746646, CER: 0.532223, loss: 152.055420
• BLEU: 0.100805

Fine-Tune small

• lm from realease 0.9.0
• 5 epochs
• tested on test07
• no further parameter changes
• n-hiddenstates 1024
• WER: 0.730518, CER: 0.511030, loss: 147.559677
• BLEU: 0.101

Transfer large

• Test on sg-speech07/test.csv
• 25 epoches
• early stopping
• lr: 0.0001, dropout 0.4, alpha and beta from paper
• WER: 0.594367, CER: 0.345339, loss: 144.413696
• BLEU: 0.236171

Fine tune Archimob

• 11 epochs
• fine tuned on transfer model above
• lr: 0.0001, dropout 0.4, alpha and beta from paper
• archimob-swissgerman-deepspeech-importer/Final_Training_CSV_for_Deepspeech_DE//test.csv
• WER: 0.679839, CER: 0.404325, loss: 58.863865
• BLEU: 0.135636

arichmoib tranfer on release 0.9.0: Test on /mnt/c/Users/debor/archimob-swissgerman-deepspeech-importer/Final_Training_CSV_for_Deepspeech_DE//test.csv - WER: 0.523529, CER: 0.269599, loss: 42.446609, BLEU: 0.272148